Weakly supervised probabilistic atlas generation through multi-atlas label fusion
First Claim
1. A method to detect anatomical region of interest (ROI) from training images having class labels to help image classification performance, the method comprising:
- (a) receiving, as input, a plurality of images, each image in the plurality of images having a class label 1≤
l≤
L and a positive threshold th between 0 and 1 for use with discriminative score maps;
(b) computing a discriminative score map for each image in the plurality of images using all remaining images as training images, where the discriminative score map for a given image comprises a spatial varying discriminative score for each image location within the given image;
(c) for each class label l, smoothing any of the discriminative score maps produced for images with the label l;
(d) producing a region of interest mask for each image in the plurality of images by thresholding its discriminative score map by th such that the produced mask has value 1 for pixels with discriminative scores greater than th and 0, otherwise; and
(e) performing image classification based on region of interest masks identified in (d).
1 Assignment
0 Petitions
Accused Products
Abstract
In many medical image classification problems, distinctive image features are often localized in certain anatomical regions. The key to efficient and accurate classification in such problems is the localization of the region of interest (ROI). To address this problem, a multi-atlas label fusion technique was developed for automatic ROI detection. Given training images with class labels, the present method infers voxel-wise scores for each image showing how distinctive each voxel is for categorizing the image. The present method for ROI segmentation and for class specific ROI patch extraction in a 2D cardiac CT body part classification application was applied and shows the effectiveness of the detected ROIs.
5 Citations
16 Claims
-
1. A method to detect anatomical region of interest (ROI) from training images having class labels to help image classification performance, the method comprising:
-
(a) receiving, as input, a plurality of images, each image in the plurality of images having a class label 1≤
l≤
L and a positive threshold th between 0 and 1 for use with discriminative score maps;(b) computing a discriminative score map for each image in the plurality of images using all remaining images as training images, where the discriminative score map for a given image comprises a spatial varying discriminative score for each image location within the given image; (c) for each class label l, smoothing any of the discriminative score maps produced for images with the label l; (d) producing a region of interest mask for each image in the plurality of images by thresholding its discriminative score map by th such that the produced mask has value 1 for pixels with discriminative scores greater than th and 0, otherwise; and (e) performing image classification based on region of interest masks identified in (d). - View Dependent Claims (2, 3, 4, 5, 6)
-
-
7. An article of manufacture having non-transitory computer readable storage medium comprising computer readable program code executable by a processor in a mobile device to implement a method to detect anatomical region of interest (ROI) from training images having class labels to help image classification performance, the non-transitory computer readable storage medium comprising:
-
(a) computer readable program code receiving, as input, a plurality of images, each image in the plurality of images having a class label 1≤
l≤
L and a positive threshold th between 0 and 1 for use with discriminative score maps;(b) computer readable program code computing a discriminative score map for each image in the plurality of images using all remaining images as training images, where the discriminative score map for a given image comprises a spatial varying discriminative score for each image location within the given image; (c) computer readable program code, for each class label l, smoothing discriminative score maps produced for images with label l; (d) computer readable program code producing a region of interest mask for each image in the plurality of images by thresholding its discriminative score map by th such that the produced mask has value 1 for pixels with discriminative scores greater than th and 0, otherwise; and (e) computer readable program code performing image classification based on region of interest masks identified in (d). - View Dependent Claims (8, 9, 10, 11, 12)
-
-
13. A method to detect anatomical region of interest (ROI) from training images having class labels to help image classification performance, the method comprising:
-
(a) receiving, as input, a plurality of images, each image in the plurality of images having a class label 1≤
l≤
L and a positive threshold th between 0 and 1 for use with discriminative score maps;(b) computing a discriminative score map for each image in the plurality of images using all remaining images as training images, where the discriminative score map for a given image comprises a spatial varying discriminative score for each image location within the given image, wherein the step of computing a discriminative score map for one target image I with label l using a set of training images I1, . . . , In comprises; i. calculating a deformable transformation between each training image Ii in the set of training images I1, . . . , In and the target image I based on an Advanced Normalization Tools (ANTs) registration algorithm; ii. warping each training image Ii in the set of training images I1, . . . , In to align with the target image I using the deformable transformation produced in step 13(b)(i), where the resulting warped image for Ii is Fi; iii. for each location x in the target image I, calculating a non-negative weight wi(x) for each warped training image Fi at location x by a joint label fusion algorithm using image intensity information in a neighborhood of x; and iv. calculating the discriminative score map for the target image at location x by summing the weights calculated in 13(b)(iii) for training images with class label l divided by the summed weights for all training images; (c) for each class label l, smoothing discriminative score maps produced for images with label l, wherein the step of smoothing discriminative score maps produced for images with label l comprises; i. receiving, as input, images with class label l, I1, K, In l , and their corresponding discriminative score maps, S1, K, Snl , and iteration number IT;ii. for each image Ii (1≤
i≤
nl), calculating deformable transformation between each of the remaining images Ij (j≠
i) and Ii using Advanced Normalization Tools (ANTs) registration algorithm;iii. for each image Ii (1≤
i≤
nl), warping each of the remaining images to Ii using the respective deformable transformation calculated in 13(c)(ii), with the resulting warped image for Ij (j≠
i) is Fj;iv. for each location x in Ii, calculating a non-negative weight wj(x) for each image Fj (j≠
i) at location x by a joint label fusion algorithm using image intensity information in a neighborhood of x;v. updating image Ii'"'"'s smoothed discriminative score at location x by - View Dependent Claims (14, 15, 16)
-
Specification